Repository navigation
Conversation
…+audio templates The vendor audit (cards, kandinskylab/kandinsky-6, the diffusers pipeline source, PR #14949) corrects the issue on Pro-distill's steps (10, not 16), the pretrain checkpoints' use, and SR's output size, and finds the pipeline defaults (512x768, guidance 5.0) are not the model's - the templates set 480x864 and guidance 1.0 explicitly. The repo is pinned to a revision because the vendor re-keys weights as diffusers' module names change. Two stored prompts in the vendor beautifier's trained shape; the compact listing budget goes to 10_250 for the new family. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…line keys Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed on huggingface_hub 1.33 snapshot_download labels the xet transfer bar with set_description_str once every file is in; dw's tqdm stand-in had no such method, so a complete download raised at the end and was marked failed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er_entry over windows Total 12.2 min (job 5d46a62ec00e, 3 windows, cuda/RTX 3090, cold); per_entry 1.56 min = one warm restore window (1.60, 1.48) plus its window_video slice. Compact catalog budget to 10_050 for the ~35 tokens. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…revert is #664 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d entry Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first mini-ai run finished its ten steps (9:06, 54 s/step) and the worker was SIGKILLed in the untiled decode of 121 frames at 480x864. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ode keeps MPS memory per tile Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts: # tests/test_catalog_structure.py
…mpt starts from the still Both templates succeeded on lem: 864x480, 121 frames at 24 fps, 44.1 kHz audio (the audio_vae rate fix, live). text-to-video 3.41 min including the first load, image-to-video 3.54 min warm, peaking near 10GB. The dragon prompt is rewritten from the vendor still itself - the image is the first frame, so the paragraph starts from what is in it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Picks text- or image-to-video, states the distilled schedule (10 steps at guidance 1.0, PiFlow refuses custom sigmas), the pipeline defaults that are not the model's (512x768, guidance 5.0), the 16 px / 4k+1 / 509-frame rules and the 44.1 kHz soundtrack, each pinned to the diffusers module that enforces it. Prompts point at Kandinsky Lab's beautifier system prompts. Stops on mps (the VAE decode is parked there) and checks for the pipeline class before trusting the diffusers version string. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… that fits 24GB Kandinsky6SRPipeline re-renders the Lite-distill clip at 2x (1728x960) in tiles on the distilled VSR checkpoint, transformer compiled (its sparse flex attention only fits compiled), and pair_audio restores the soundtrack. At the default tiles the SR VAE's per-tile decode needs about 26GB: the 512x768 tile's 6.38GiB fp32 normalize copy on top of 19.7GiB. Holding the transformer's tile_sizes to the trained 512x512 at scale 2 peaks at 19.3GiB. Run end to end on lem's RTX 3090 in 9.17 min; the upscale adds real texture over a bicubic 2x. The skill offers it on CUDA and pins both limits; the compact budget entry covers all three templates at 10_450. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the download status fix for the next release Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…job ae52011e2f05) Lead, tail and delivered rate now come from re-transcribed probe-1 clips, checkable over MCP. Tail 0.09-0.88 s, median about 0.37 s. The envelope figures stay as a cross-check, and clipping is still refuted. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…hots resolution, speech_in_silent_shot, per-shot clipping Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the map row Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…default length from diffusers Arch-review bounce: _render_length rounds through aligned() and snap_guide_length's 17m+5 tail through aligned_down() on one RENDER_GRID declared beside the chunk constants; RENDER_BASE, the hand-rolled ceil and DEFAULT_NUM_FRAMES are gone - the default is the stock layout step's own num_frames InputParam default. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rows 9-13 of the H3 dialogue-probe audit: extras (inconclusive, keep 3), off-screen naming (confirmed) and negated naming (the negation lint stays cut), voice words in a silent sound line (no invented speech), a quoted sign (legible, no enhancer issue; a blank sign invents text), and Music3 instrumentals (inconclusive: Whisper hallucinates on music). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…elds A job stores the card it ran on as device_ordinal and device_card, set at dispatch; job.device is label_of() joining the two. History gains both columns behind a PRAGMA-guarded ALTER with a one-time SQL backfill that splits the stored label. Rerun affinity, workers() and observed cost read the fields; card_of and ordinal_of are gone. REST/MCP `device` unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ign record Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ss threads Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…iterals via whole_number
The validate-time window-count check skipped num_frames/overlap given as a
numeric string ("33.0"), so validate passed a list the run refuses (C-F354h).
_literal_whole now reads through task_domains.whole_number, the run's rule.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…each cut, not the whole song from sample 0 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…n H3 templates Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…in with an AST test Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
`host_memory_fields` reads /proc/self/status for the anon/file split (#709), which the test never stubbed, so it passed on macOS and failed on Linux CI. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… cycle task_domains imported check_script's parsers from the task, and the task imports task_domains (through assess, audio_utils and joins) - a 13-module cycle the import_cycles ratchet has reported on every develop push since f1e0d48. The parsers (parse_lines, parse_shots, shot_names_error, with strip_markup and normalize_words) move to dw/script_lines.py, which needs only dw.shots; both the task and validation call the same functions, as the map row requires. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…for 0.10.0 The ratchet has been red on every develop push since f1e0d48 (2026-10-05), masked in CI by a test failure in the step before it. The import cycle is fixed (2e536f7); the rest is raised here so 0.10.0 can ship, and #790 lists every offender to bring back under the old baseline: - modules 175 -> 198: the new tasks and pipelines, the jobs.py (#776) and h3_blocks (#771) splits, dw/outbound.py (#690), dw/script_lines.py - modules_over_size_ceiling 0 -> 2: dw/task_domains.py, dw/tasks/task.py - functions_over_150_lines 0 -> 7: run_chain, load_component, _make_guide_blocks, _make_blocks, concat_videos, plan_cuts, check - complex_functions 7 -> 16: the H3 guide and hold block builders, locations._walk, lut.parse_cube, admission.argument_reference_errors, script_check.check and the rest named in #790 - duplicate_blocks 5 -> 8: type_references/reference_names/hold_audio/ guides/kernel_availability/step_value_checks/reference_limits/ introspection/task_domains pairs - test_dw_patch_targets 284 -> 325, claude_md_lines 128 -> 129 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Merge develop into master for 0.10.0. Preflight passed on the Mac (MPS): 13,342 unit, integration, 109 e2e, architecture ratchet green after 2e536f7 (import cycle fixed) and 321cdbe (baseline raised; offenders in #790).
Release notes are drafted under 0.10.0 in docs/RELEASING.md.
🤖 Generated with Claude Code